Knowledge Graphs and Their Applications in Meituan's AI Platform

Knowledge graphs are a key knowledge representation method in the AI era, breaking data silos across scenarios and supporting search, recommendation, QA, explanation, and decision‑making. Meituan’s AI platform has built a massive knowledge graph for the life‑entertainment domain to link users and merchants comprehensively.

In a recent online talk, Dr. Wang Zhongyuan, head of Meituan’s NLP Center, presented the design, construction, challenges, and practical applications of the "Meituan Brain" knowledge graph.

The talk highlighted the rapid impact of AI products such as AlphaGo, Amazon Go, Skype Translator, and Siri, emphasizing that their success relies on advances in machine learning, computer vision, speech recognition, and natural language processing.

Meituan aims to enable machines to read and fully understand user reviews, summarizing merchant information for users and eventually providing an AI assistant for dining and entertainment decisions.

Two major technical drivers for AI are deep learning and knowledge graphs. Deep learning excels at specific tasks but requires massive data, high compute, and offers limited interpretability. Knowledge graphs, by contrast, provide reusable, highly interpretable knowledge across tasks.

Knowledge graphs can be categorized into Common Sense Knowledge Graphs (e.g., WordNet, Probase, NELL) and Encyclopedia Knowledge Graphs (e.g., Google Knowledge Graph, Microsoft Concept Graph, Meituan Brain). The former focuses on linguistic relations, while the latter emphasizes entities and factual relations.

Meituan’s "Common Sense Knowledge Graph" (Probase) contains millions of nodes (concepts, entities, attributes, verbs, adjectives) and edges such as isA and isPropertyOf. Conceptualization models map short texts to millions of concepts, supporting tasks like NER, text tagging, ad matching, and query recommendation.

The "Encyclopedia Knowledge Graph"—Meituan Brain—covers billions of user reviews, tens of thousands of merchant tags, and over 140 million dishes. It is built using statistical language models, topic models (LDA), and deep learning models (LSTM, Bi‑LSTM, CRF, CNN+LSTM) to extract merchant tags, dish attributes, and fine‑grained sentiment.

For sentiment analysis, a CNN+LSTM architecture predicts overall sentiment and fine‑grained dimensions (traffic, environment, hygiene, taste, etc.), a capability that is still rare worldwide.

Applications of the knowledge graph include improved search ranking (e.g., surfacing restaurants serving spicy crayfish based on sentiment), personalized recommendations in shopping districts, and AI assistants that answer user queries using the graph.

The system relies on large‑scale graph storage and computation engines, GPU clusters for deep learning, and advanced models such as a "water‑wave" deep learning model presented at CIKM.